Request
Add ZCode (Z.ai's coding agent harness, https://zcode.z.ai) as a first-class --target for codegraph install, so codegraph install --target=zcode writes the MCP config automatically.
Related: tracking issue #649 (zcode is already listed there from #962). Opening a dedicated issue to attach the concrete config format and hook details so the target is easier to land.
Why
ZCode is a first-class MCP + hooks-capable agent. Today, ZCode users have to hand-edit the MCP config. CodeGraph works great once configured manually — the ask is just the one-command install target like the other agents have.
What ZCode needs
1. MCP server config (install target)
ZCode reads MCP servers from ~/.zcode/cli/config.json (user scope) or <workspace>/.zcode/config.json (workspace scope), under mcp.servers:
{
"mcp": {
"servers": {
"codegraph": {
"type": "stdio",
"command": "codegraph",
"args": ["serve", "--mcp"],
"enabled": true,
"timeoutMs": 60000
}
}
}
}
That's it for MCP — same codegraph serve --mcp as everywhere else. A new file in src/installer/targets/ plus a registry entry, as you noted in #649.
2. (Bonus) UserPromptSubmit prompt-hook
ZCode also exposes a UserPromptSubmit hook that can inject context per turn. If CodeGraph shipped a prompt-hook for ZCode (analogous to what exists for Claude Code), it could inject relevant-symbols context on each prompt. This is optional — the MCP target alone is the main ask.
ZCode hooks docs: https://zcode.z.ai/cn/docs/hooks
ZCode plugin/MCP docs: https://zcode.z.ai/cn/docs/plugin
Config location summary
| Scope |
Path |
| User |
~/.zcode/cli/config.json |
| Workspace |
<workspace>/.zcode/config.json |
Happy to test a PR on Windows if that helps.
Request
Add ZCode (Z.ai's coding agent harness, https://zcode.z.ai) as a first-class
--targetforcodegraph install, socodegraph install --target=zcodewrites the MCP config automatically.Related: tracking issue #649 (zcode is already listed there from #962). Opening a dedicated issue to attach the concrete config format and hook details so the target is easier to land.
Why
ZCode is a first-class MCP + hooks-capable agent. Today, ZCode users have to hand-edit the MCP config. CodeGraph works great once configured manually — the ask is just the one-command install target like the other agents have.
What ZCode needs
1. MCP server config (install target)
ZCode reads MCP servers from
~/.zcode/cli/config.json(user scope) or<workspace>/.zcode/config.json(workspace scope), undermcp.servers:{ "mcp": { "servers": { "codegraph": { "type": "stdio", "command": "codegraph", "args": ["serve", "--mcp"], "enabled": true, "timeoutMs": 60000 } } } }That's it for MCP — same
codegraph serve --mcpas everywhere else. A new file insrc/installer/targets/plus a registry entry, as you noted in #649.2. (Bonus) UserPromptSubmit prompt-hook
ZCode also exposes a
UserPromptSubmithook that can inject context per turn. If CodeGraph shipped a prompt-hook for ZCode (analogous to what exists for Claude Code), it could inject relevant-symbols context on each prompt. This is optional — the MCP target alone is the main ask.ZCode hooks docs: https://zcode.z.ai/cn/docs/hooks
ZCode plugin/MCP docs: https://zcode.z.ai/cn/docs/plugin
Config location summary
~/.zcode/cli/config.json<workspace>/.zcode/config.jsonHappy to test a PR on Windows if that helps.